Mask-based Text Scoring for Product Title Summarization

Abstract

In e-commerce, long product titles with rich information help attract users, but they are usually truncated for display on small-screen mobile devices, which results in neglection of important information and in turn low click-through rate. This paper presents a novel product title summarization method via the use of a mask-based text information scoring network. Via quantified evaluation of expressiveness, the most telling points are identified from the original title for a concise version which best retains its content. Our experiments show that, even without external information, our proposed method MPTS outperforms established benchmark models by 1.48% (ROUGE-1), 5.11% (ROUGE-2) and 1.37% (ROUGE-L) respectively.